AI Engineer's Stop Telling Agents How — Dan Barrett, Raymond Weitekamp & Jose Montes de Oca, OpenProse: skim's analysis identifies 10 key moments, with 1 potential conflict of interest flagged. OpenProse creators Dan Barrett and Raymond Weitekamp demonstrate how to write multi-agent workflows as declarative markdown contracts that any coding harness can compile and execute.
Category: Tech. Format: Educational. YouTube video analyzed by skim.
skim AI Analysis
Credibility assessment: Creator Demo & Technical Pitch. Dan Barrett and Raymond Weitekamp present live software demonstrations of their own open-source framework, OpenProse, showing actual pull request logs and Codex terminal execution. However, their claims regarding enterprise banking cost savings and client adoption rely on self-reported experience without independent documentation.
Bias assessment: Product-Evangelist Framing. As the creators of OpenProse and operators of its commercial cloud platform, Dan Barrett and Raymond Weitekamp naturally frame declarative markdown contracts as the superior next evolution in AI engineering, though they candidly discuss current verification weaknesses and cases where standard chat is better.
Originality: 86% — Novel Meta-Harness Design. Treating the language model itself as the compiler and virtual machine for a typed markdown specification offers a distinctive alternative to deterministic Python frameworks like DSPy. The live demonstrations of webhook-triggered adversarial PR loops and quant documentation illustrate practical, original engineering.
Depth: 77% — Architectural & Practical. The speakers go beyond a surface pitch to examine trade-offs between declarative outcomes and procedural pinning, the self-review bias of single-session coding agents, comparisons with DSPy, and where non-generative decision classifiers fit into fuzzy agent execution.
Key Points (10)
1. Dan Barrett on Declaring Outcomes Over Procedures
Timestamp: 00:02:12 to 00:05:41
Dan Barrett traces AI engineering from token prompting and manual context stuffing to self-modifying harnesses that manage sub-agents automatically. Because modern harnesses handle execution, engineers should write composable, declarative contracts stating what outcome they want rather than how to get it. Dictating procedural steps wastes model intelligence and prevents cheaper models from executing the workflow.
Significance (High): Shifting from procedural instructions to declarative contracts treats prompts like software specifications, letting teams upgrade underlying agent harnesses without rewriting their core operational logic.
Sources in support: Dan Barrett (Creator of OpenProse)
2. Raymond Weitekamp Explains Pure-Markdown Workflows
Timestamp: 00:05:55 to 00:09:27
Raymond Weitekamp defines OpenProse around two pillars: workflows are written entirely as plain markdown text, and execution is event-driven rather than chat-based. Because both the workflow and the runtime compiler are just text prompts, regulated enterprises like banks can adopt multi-agent orchestration on existing coding agents without new procurement or IT deployments.
Significance (High): Packaging an entire multi-agent framework and its compiler as plain markdown files bypasses enterprise software procurement bottlenecks while keeping workflows readable for non-engineers.
Sources in support: Raymond Weitekamp (OpenProse Engineer and Creator of run.prose.md)
3. Wiring Sub-Agents, Tools, and Event Triggers
Timestamp: 00:09:58 to 00:14:51
OpenProse contracts let users declare sub-agents, required CLI skills, specific tools like Exa search, and strict input-output requirements that halt compilation if dependencies are missing. Instead of relying on interactive chat, the hosted run.prose.md platform pairs these contracts with cron schedules or webhooks from services like Slack and GitHub.
Significance (High): Enforcing strict tool dependencies and attaching contracts to webhooks turns ad-hoc coding agents into standing background services that react automatically to repository events.
Sources in support: Raymond Weitekamp (OpenProse Engineer and Creator of run.prose.md)
4. Pinning Winning Agent Traces to Cut Compute Costs
Timestamp: 00:15:16 to 00:18:59
Responding to an audience question, Dan Barrett and Raymond Weitekamp explain that users can start with a loose declarative goal solved by an expensive frontier model and then pin that execution trace into a reusable procedural workflow. Raymond Weitekamp adds that beneath a user's short markdown file sits a system prompt of roughly 70,000 to 80,000 tokens that overrides markdown headers with a strict type system.
Significance (High): Capturing a successful frontier-model trace and pinning it into a procedural contract lets teams shift recurring workloads onto cheaper models without sacrificing reliability.
Sources in support: Dan Barrett (Creator of OpenProse), Raymond Weitekamp (OpenProse Engineer and Creator of run.prose.md)
5. Live Demo of Voice-Created Adversarial PR Reviews
Timestamp: 00:21:20 to 00:25:16
Raymond Weitekamp demonstrates run.prose.md by asking a voice agent to build a contract that reviews every new pull request like a grumpy senior developer skeptical of AI code. While the agent generates the markdown program and awaits manual confirmation to spend money, he shows how his own repository uses webhook-triggered review agents as automated backpressure.
Significance (High): Benchmarking every pull request in a fresh agent session against a written contract eliminates the self-grading bias common when a coding agent reviews its own output.
Sources in support: Raymond Weitekamp (OpenProse Engineer and Creator of run.prose.md), Dan Barrett (Creator of OpenProse)
6. Contrasting OpenProse With DSPy and Verifiers
Timestamp: 00:25:18 to 00:29:18
Addressing an audience question about one-shot decision classifiers, Raymond Weitekamp compares OpenProse to DSPy, noting both provide abstraction layers for complex workflows but take opposite implementation paths. While DSPy isolates every language model call inside deterministic code and typed signatures, OpenProse lets the AI act as a fuzzy compiler and decide when a contract is met.
Significance (Medium): Pairing fuzzy LLM compilation with fast, deterministic verifier models addresses OpenProse's main vulnerability: agents prematurely declaring their own work complete.
Sources in support: Raymond Weitekamp (OpenProse Engineer and Creator of run.prose.md), Dan Barrett (Creator of OpenProse)
7. On-Premises Quant Model Documentation for Banks
Timestamp: 00:31:21 to 00:38:15
Dan Barrett explains that enterprises can run OpenProse entirely on-premises over open models and internal harnesses so they never ship proprietary data or standard operating procedures to external AI labs. He demonstrates a workflow in Codex that boots the public OpenProse markdown kernel to generate regulatory documentation for a quantitative finance model.
Significance (High): Treating regulatory compliance checklists as executable markdown contracts lets financial institutions automate millions of dollars in recurring quant documentation completely on-premises.
Sources in support: Dan Barrett (Creator of OpenProse), Raymond Weitekamp (OpenProse Engineer and Creator of run.prose.md)
8. When to Use Exploratory Chat vs Repeatable Prose
Timestamp: 00:38:28 to 00:41:54
When asked how users formulate clear contracts, Dan Barrett notes that the OpenProse kernel includes best-practice and anti-pattern guides that helper agents use to draft contracts iteratively. Raymond Weitekamp observes that chat interfaces succeeded because users can explore vague ideas without knowing what they want upfront, whereas frameworks like DSPy demand rigid return signatures immediately.
Significance (Medium): Drawing a clear boundary between exploratory chat and repeatable contract execution helps teams avoid over-engineering one-off queries while standardizing recurring operations.
Sources in support: Dan Barrett (Creator of OpenProse), Raymond Weitekamp (OpenProse Engineer and Creator of run.prose.md)
9. Composing Contracts Into Multi-Agent Graphs
Timestamp: 00:41:57 to 00:47:02
Dan Barrett explains that OpenProse supports injecting multiple contracts into a single run and chaining upstream and downstream dependencies like React memoization or a directed acyclic graph. When upstream data changes, semantic classifiers can detect staleness and trigger downstream contracts as intermediary checkpoints.
Significance (High): Composing intermediary contracts into directed graphs gives engineers checkpoint control over multi-hour agent runs without hard-coding imperative orchestration logic.
Sources in support: Dan Barrett (Creator of OpenProse), Raymond Weitekamp (OpenProse Engineer and Creator of run.prose.md)
10. Evolving the Kernel and Adoption Beyond Coding
Timestamp: 00:47:06 to 00:49:23
Dan Barrett shares that the OpenProse kernel is refined using empirical benchmarks and an evolutionary algorithm that shrinks the system prompt while testing whether agents still conform to contracts. Raymond Weitekamp illustrates skill chaining by describing an automated Matt Pocock workflow that forces sub-agents to load sequential questioning and test-driven development skills.
Significance (Medium): Shrinking the kernel via evolutionary benchmarking lowers token overhead, while markdown's readability expands multi-agent automation to operations and product teams.
Sources in support: Dan Barrett (Creator of OpenProse), Raymond Weitekamp (OpenProse Engineer and Creator of run.prose.md)
This analysis was generated by skim (skim.plus), an AI-powered content analysis platform by Credible AI. Scores and classifications represent the platform's AI-generated assessment and should be considered alongside other sources.
